Breast Cancer Detection in Digital Mammograms

Transcription

1 Breast Cancer Detection in Digital Mammograms Kanchan Lata Kashyap *, Manish Kumar Bajpai, Pritee Khanna Computer Science & Engineering, Indian Institute of Information Technology, Design & Manufacturing Jabalpur, India *Corresponding Author Abstract This paper discusses an approach for automatic detection of abnormalities in the mammograms. Image processing techniques have been applied to accurately segment the suspicious region-of-interest (ROI) prior to abnormality detection. Unsharp masking has been applied for enhancement of the mammogram. Noise removal has been done by using median filtering. Discrete wavelet transform has been applied on filtered image to get the accurate result prior to segmentation. Suspicious ROI has been segmented using the fuzzy-c-means with thresholding technique. Tamura features, shape based features and moment invariants are extracted from the segmented ROI to detect the abnormalities in the mammograms. Proposed algorithm has been validated on the Mini-MIAS data set. Keywords Unsharp masking, Median filtering, fuzzy-cmeans, discrete wavelet transform I. INTRODUCTION Mammography has been considered as the most important technique to investigate the breast cancer. It can be used to detect the disease in the early stage when recovery is possible. Aim of Computer Aided Diagnosis system (CAD) is to read the mammograms, locate the suspicious regions of abnormalities and analyze its characteristics. Performance and reliability of CAD depends on accurate segmentation of the lesions and appropriate feature selection. Global segmentation of the lesions is the most challenging task due to the artifacts and healthy tissues present in the mammograms [1]. Various algorithms have been developed in literature for early detection of the breast cancer in mammograms. Kegelmeyer et al. presented laws textural feature and binary decision tree classifier to classify between lesions and normal tissues [2]. Campanini et al. presented an SVM classifier for mass detection in digital mammograms [3]. Sahiner et al. used four gray-level difference statistics (GLDS) texture features and convolution neural network for mass detection [4]. Bellotti et al. proposed an automated CAD system for mass detection based on gray level co-occurrences matrices (GLCM) features and back-propagation neural network classifier [5]. Mudigonda et al. and Dominguez et al. proposed Density slicing technique for mass segmentation [6, 7]. Sharma and Khanna developed a CAD system to detect the breast cancer using Zernike moments [8]. Kashyap et al. proposed an efficient segmentation algorithm for breast cancer detection in mammograms [9]. Present work proposes a shape based approach for breast cancer detection using mammograms. Organization of the paper is given as follows: Flow chart and the proposed algorithm is described in Section 2. Section 3 includes the validation of algorithm and experimental results followed by conclusions in Section 4. II. PROPOSED ALGORITHM Flow chart of the proposed algorithm is given in Fig 1 which includes breast region segmentation, enhancement, filtering, segmentation, feature extraction and classification steps. Different types of noises such as tape artifacts, labels or scanning artifacts are present in the mammogram images. Such noises must be suppressed and excluded from further processing. Breast region is segmented after removal of noises which contains the useful information. Original mammogram images are binarized by global thresholding technique. Morphological opening operation is performed to remove small regions [10]. Connected component labeling algorithm is applied to identify the largest connected component as a breast region. Breast region is superimposed on the original image to get the final segmented unlabeled image I P. Contrast of Mammogram images are enhanced by using unsharp masking technique to get enhanced image I PE [10]. Original unlabeled image is inverted to get inverted ' images I. Inverted image is also enhanced by using p unsharp masking technique and it results enhanced inverted image I '. Enhanced inverted image is subtracted from PE Enhanced image I PE and which results in subtracted image I 1. Subtracted image is filtered by applying median filter to remove the overall noise of image and it results filtered image I 1F. Discrete wavelet transform (DWT) is applied to decompose the filtered image into approximated image and detailed image [10]. Wavelet transformation provides both the frequency as well as temporal information. Wavelet transform is a tool which has been used to locate the low frequency area and high frequency area in image I 1F. It also highlights the structural, geometrical and directional features of mammographic image. Detailed image is reconstructed and added to filtered image to sharpen the edges of mass. Sharpen image is used further to segment the suspicious mass region. Segmentation is performed to partition the images into homogeneous regions and extract the region-of-interest (ROI). Fuzzy C-means (FCM) clustering with thresholding algorithm is applied on sharpened image [11]. Connected component algorithm is applied to extract the segmented suspicious mass region. A set of tamura features, shape based and moment invariant features are extracted from each segmented suspicious regions and feature vector is created. Support vector machine (SVM) is used to classify the segmented region as an abnormal region or normal region [12] /15/$ IEEE

2 Inverted Image I ' P Enhancement ' I PE Original Image I Breast Region Segmentation I P Enhancement I PE Subtraction ' I = I PE I 1 PE Filtered Image I 1F Decomposition I 1HH Isharpend= I1 F + I1 HH Fig. 1. Flow chart of proposed algorithm Segmentation I Seg Feature Extraction Classification Abnormal region Normal region Algorithm of the proposed methodology is given as follows: Algorithm1 ( I, Class) Input : I I : input Image I P : preprocessed image I PE : enhanced preprocessed image I P : inverted preprocessed image I PE : enhanced Inverted preprocessed image I 1 : subtracted Image I 1F : filtered Image I 1FA : approximated decomposed filtered image I 1FD : detailed decomposed filtered image I seg : segmented image FV : feature vector Preprocessing ( ) : preprocessing of image Unsharp ( ) : standard unsharp masking mthod Invert ( ): standard inversion method of image Median ( ) : standard median filter Wavelet ( ): standard wavelet method for image decomposition FCMthreshold ( ) : fuzzy c means clustering with thresholding Feature_vector ( ): function to extract features SVM ( ): standard function for classification Output : Class begin end I P = Preprocessing (I) I PE =Unsharp(I P ) I P =Invert(I P ) I PE=Unsharp(I P) I 1 = I PE - I PE I 1F = Median(I 1 ) [I 1FA, I 1FD ] = Wavelet( I 1F ) I Sharpened = I 1F + I 1FD I Seg = FCMthreshold( I Sharpened ) FV = Feature_vector(I seg ) Class= SVM(FV) III. EXPERIMENTAL RESULTS AND DISCUSSION Proposed algorithm is validated on publicly available Mammographic Image Analysis Society (MIAS) data set containing 322 mammogram images available in this data set. All sample images are scaled down to pixels

3 using nearest neighbor interpolation method. Removal of tape artifacts and labels are done in first stage of proposed algorithm. Fig. 2 shows the intermediate results of breast region segmentation. Fig 2(a), (b), (c) and (d) show the original mammogram, binarized image after performing thresholding operation, image after performing morphological opening operation and unlabeled image, respectively. It is clear from Fig. 2 that background information is successfully removed from the mammographic image. (a) (b) Fig. 4. (a) Inverted unlabeled image (b) Enhnaced inverted image (a) Original image (b) Binarized image (a) (b ) Fig. 5. (a) Subtracted image (b) Filtered image (c) Morphological opened image (d) Unlabeled image Fig. 2. Intermediate results of breast area identification Fig. 3 shows the enhanced unlabeled image. Inverted unlabeled mammogram and enhanced inverted mammogram images are shown in Fig. 4. Subtracted image is shown in fig 5. It is clear that suspicious mass region is enhanced in the subtracted images. Filtered image and segmented suspicious mass regions are shown in Fig. 6. The sample segmented mammogram images after applying the FCM algorithm with thresholding technique are shown in Fig. 7. Connected component labeling algorithm is applied to extract the segmented suspicious mass. Fractal dimension using box counting method, three tamura based features i.e., coarseness, contrast and directionality along with ten region based features i.e., area, perimeter, Fig. 3. Enhanced unlabeled image (a) Sharpened image (b) Segmented image Fig. 6. Results of suspicious region segmentation major axis length, minor axis length, eccentricity, orientation, convex area, equi- diameter, solidity, extent are extracted from segmented region. Seven moment invariant are also extracted from the segmented suspicious mass region [10, 13, 14, 15]. Fisher score technique has been used to select the prominent features from the all extracted features [12]. Three region based features eccentricity, solidity and extent are selected from nine features and the first three features are selected from seven moment based features. Support vector machine with radial basis function (RBF) kernel is used to classify. Performance of the classification is measured in terms of sensitivity, specificity, accuracy and ROC curve [16]. Sensitivity measures the proportion of actual positives which are correctly identified as cancerous. Specificity measures the proportion of actual negative which are correctly identified as noncancerous.

4 Mathematical expression of sensitivity, specificity and accuracy are given as follows: Mdb28 Sensitivit y TP = TP + FN TN Specificity = FP + TN (1) (2) Mdb132 Accuracy TP + TN = (3) TP + FN + FP + TN here TP, TN, FP and FN represents the true positive, true negative, false positive and false negative, respectively. ROC curve is a plot between the true positive rate (TPR) and false positive rate (FPR). Performance measurement of proposed algorithm is given in Table 1. TABLE I PERFORMANCE MEASURE OF PROPOSED ALGORITHM Mdb134 Features Sensitivity Specificity Accuracy Region based features 100% 83.34% 91% Momentinvariant +fractal dimension 97.14% 96.67% 96.92% Tamura features 85.71% 70% 78.46% Mdb184 Mdb271 Mdb81 (a) Sharpened image ( b) Segmented image Fig. 7. Results of suspicious region segmentation

5 IV. CONCLUSIONS Proposed method presents an algorithm to enhance, segment and classify abnormalities present in the mammogram. The following conclusions can be derived from the work: Figure-8 ROC curve for abnormality classification 100%, 97.14%, 85.17% sensitivity and 83.34%, 96.67%, 70% specificity has been achieved using region based features, combination of moments with fractal dimension and tamura based features, respectively as shown in the table 1. Area under curve A z of 0.91, 0.96, 0.78 has been achieved as shown in the fig 8. Performance of the proposed algorithm with other existing methods in this field is summarized in Table-2. TABLE II COMPARISIONS WITH EXISTING METHODS Methods Features Classifier No. of images Proposed Algorithm Brzakovic et al[17] Cao et al.[18] Pasquale et al.[19] Region based, Moment invariant +fractal dimension, Tamura features Shape based features Intensity, shape and texture Intensity, shape, texture Accuracy SVM % Decision tree 96.92% 78.46% 25 85% SVM % Sensitivity Feed Forward Neural Network % Sensitivity Subtraction of enhanced preprocessed and enhanced inverted preprocessed image improves the detection of suspicious region in mammograms. Edges of the suspicious region are sharpened by adding detailed coefficients of the wavelet decomposition with filtered image. Accuracy of the segmentation is improved by using FCM algorithm. Moment based features has given better result than the region based and tamura based features. Accuracy of the proposed algorithm will be improved by extracting some other features of suspicious region. The Classification of the benign and malignant mass will be done in further research work. REFERENCES [1] Sharma, P. Khanna, ROI Segmentation using Local Binary Image, IEEE International conference on System, Computing and Engineering, pp , [2] W. P. Jr Kegelmeyer, J. M. Pruneda, P. D. Bourland, A. Hillis, M. W. Riggs, and M. L. Nipper, Computer-aided mammographic screening for speculated lesions, Radiology, vol. 191, no. 2, pp , [3] R. Campanini, D. Dongiovanni, E. Iampieri, N. Lanconelli, M. Masotti, G. Palermo, A. Riccardi, and M. Roffilli, A novel featureless approach to mass detection in digital mammograms based on support vector machines, Phys. Med. Biol., vol. 49, no. 6, pp , [4] B. Sahiner, H. P. Chan, N. Petrick, D. Wei, M. A. Helvie, D. D. Adler, and M. M. Goodsitt, Classification of mass and normal breast tissue: A convolution neural network classifier with spatial domain and texture images, IEEE Trans. Med. Imag., vol. no. 15, pp , [5] R. Bellotti, F. D. Carlo, S. Tangaro, G. Gargano, G. Maggipinto, M. Castellano, R. Massafra, D. Cascio, F. Fauci, R. Magro, G. Raso, A. Lauria, G. Forni, S. Bagnasco, P. Cerello, E. Zanon, S. C. Cheran, E. L. Torres, U. Bottigli, G. L. Masala, P. Oliva, A. Retico, M. E. Fantacci, R. Cataldo, I. D. Mitri, and G. D. Nunzio, A completely automated CAD system for mass detection in a large mammographic database, Med. Phys., vol. no. 33, pp , 2006.J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp [6] N. R. Mudigonda, R. M. Rangayyan and J. E. Desautels, Detection of breast masses in mammograms by density slicing and texture flowfield analysis, IEEE Trans. Med. Imaging. 20, pp , [7] A. R. Dominguez, A. K. Nandi, Detection of masses in mammograms via statistically based enhancement multilevel-thresholding segmentation, and region selection, Comput. Med. Imaging Graphics. 32, pp , [8] S. Sharma and P. Khanna, Computer-Aided Diagnosis of Malignant Mammograms using Zernike Moments and SVM, Journal of Digital Imaging, vol. 28, pp 77-90, [9] K. L. Kashyap, M. K. Bajpai, P. Khanna, A Novel Approach for Segmentation of Mammographic image to detect breast cancer, 7th International Symposium on Process Tomography Dresden, Germany, 2015 [Accepted].

6 [10] R. C. Gonzalez., Digital image processing, 2nd Edition. India: Prentice Hall, [11] J. C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum Press, New York, [12] R. O. Duda, P. E. Hart, D. G. Stork, Pattern Classification, second ed., John Wiley and Sons, New York [13] H. Tamura, S Mori, T. Yamavaki Textural features corresponding to visual perception, IEEE Trans On Systems, Man and Cybernetics, vol 8, pp , [14] M. K. Hu, Visual pattern recognition by moment invariants, IRE Transactions on Information Theory, vol. 8, pp , [15] ] T. M. Nguyen and R. M. Rangayyan, Shape Analysis of Breast Masses in Mammograms via the Fractal Dimension, in Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, [16] J. A. Hanley and B. J. McNeil, The meaning and use of the area under a receiver operating characteristic _ROC_ curve, Radiology 143, pp , [17] D. Brzakovic,X. M. Lau, P. Brzakovic, An Approach to Automated Detection of Tumors in Mammograms, IEEE Trans. On Med. Imaging, vol.9, pp , [18] A. Cao, Q. Song, X. Yang, Robust information clustering incorporating spatial information for breast mass detection in digitized mammograms, Computer Vision and Image Understanding, 109, pp , [19] P. Delogu, M. E. Fantacci, P. Kasae, A. Retico, Characterization of mammographic masses using a gradient-based segmentation algorithm and a neural classifier, Computers in Biology using a gradient-based segmentation algorithm and a neural classifier, Computers in Biology and Medicine 37, pp [20] J. Suckling, the mammographic image analysis society digital mammogram database, in 2 nd international workshop on digital mammography, 1994.